Gas diffusion layers (GDLs) are responsible for oxygen delivery to the catalyst layer and water management in polymer electrolyte fuel cells (PEFCs). With x-ray computed tomography (CT) – a non-destructive technique, one can observe water transport through the GDL, and also assess local wettability. With ability of X-ray CT to capture dynamics of water transport the quantity of data and the time required to process the data increases. There is a need to further automate data processing and analysis through machine learning; in particular, convolutional neural networks (CNNs). CNNs can be applied to a multitude of image processing tasks such as finding centers of rotation for CT data [1], removal of reconstruction artifacts [1], and denoising of data sets [2]. However, CNNs are probably most known for their ability to perform segmentation/labeling.Although CNNs have proven to be widely applicable to image processing and analysis, it has also become apparent that individual pairings of architecture and application must be validated [1]. As such, we focus on the use of CNNs to segment synchrotron X-ray CT data sets for water within various GDLs. The water is tracked for an in-situ evaporation/condensation experiment. Knowledge of the water's location and shape as a function of time is necessary to understand local GDL’s wettability and its effect on PEFC performance. The challenge these samples present to more traditional segmentation methods is the low contrast between water and carbon fibers. Depending on the composition of a GDL, binder, polytetrafluoroethylene (PTFE), and the microporous layer (MPL) may also be present in the same narrow window of gray-scale values. Separating these phases is essential in evaluating the material's water transport properties. We will present results from both our efforts with CNNs and the specific PEFC experiment that produced the data sets.
Read full abstract7-days of FREE Audio papers, translation & more with Prime
7-days of FREE Prime access